Improving Stress Detection Using Doddington’s Menagerie

This is the authors' abstract. We don't add key points for this paper.

Objectives: This study’s objective is to improving stress detection and prediction systems using biometric categorization known as Doddington’s menagerie be user categorization. Using this categorization allows for improving the accuracy of the proposed attention-based deep learning classifier that accounts for individual variability in humans’ biometric indicators of stress. These objectives are motivated by the limitations of one-size-fits-all models, which may perform well for most subjects but yield unreliable results for specific subject categories. Materials and Methods: An open-source Stress-Predict Dataset was used, which contains physiological data from 31 participants recorded during cognitive tasks and rest periods. We applied a biometric taxonomy grounded in Doddington’s menagerie, identifying which categories frequently produce false positives and which ones account for most false negatives. Three classifiers were trained to detect the state of ‘Stress’ or ‘Rest’: a standard Long Short-Term Memory (LSTM), a Temporal Convolutional Network, and the proposed Attention-Based LSTM Fusion Network. Leave-One-Subject-Out Cross-Validation was used to evaluate the classifiers’ performance, and temporal attention-weight matrices were used for qualitative interpretability of the physiological signals. Results: The novel taxonomic analysis revealed inter-individual variability, which allowed the development of user categorization to improve classification model that is robust and responsive to participants’ diversity. The proposed Attention-Based LSTM Fusion Network demonstrated a significant performance divergence compared directly to TCN across test accuracy ( $p = 0.0261$ ), test F1-score ( $p = 0.0027$ ) and Transformer across test accuracy ( $p = 0.0017$ ), test F1-score ( $p = 0.0008$ ), achieving a mean test accuracy of 0.7628~± ~0.0975 and F1-score of 0.7617~± ~0.0997 , while performance after used categorization identified by Doddington’s menagerie produced a mean test accuracy of 0.7966~± ~0.0698 and an F1-score of 0.8029~± ~0.0678 .